AXIOMINFERENCE
Research
Small models.
Real adaptation.
Three ways to get more from a small model.
Train a new architecture from scratch, post-train a strong model for agents, or tune a tiny reasoner until it flies on old hardware. Each has its own page with the full story, benchmarks and a live feed from its repository.
Project Kestrel
A small, dense, adaptively-learning architecture. Kestrel-Nano (157M) is trained; Phase D and beyond are tracked on a live roadmap read straight from the repo.
Fable-Coder
Four iterations of a 7B code model for agentic tool use and security — including the failures — plus OmniAgent-Bench, a benchmark that grades by what actually works.
MalxLabs-V4
A 1.5B DeepSeek-R1 distill, fine-tuned in the cloud and tuned to run at 80+ tokens/s on a GTX 1080, with real head-to-head benchmarks against its base.
How they fit together.
The same question — how much capability can you get on hardware you own? — approached from three directions.
| Program | Approach | Starting point | Trained on | Runs on |
|---|---|---|---|---|
| Project Kestrel | Design and train a new architecture: hybrid attention, looped core, sparse memory | Nothing — ~157M parameters from scratch | One rented RTX 3090 for v0.1, now a local RTX 3060 | Nearly any computer |
| Fable-Coder | Replay-anchored LoRA post-training for agentic tool use and security | Qwen2.5-Coder-7B-Instruct | A100 (V4 in ~32 minutes) | An RTX 3060 (4.36 GiB GGUF) |
| MalxLabs-V4 | Supervised fine-tune for math and code reasoning, with aggressive inference tuning | DeepSeek-R1-Distill-Qwen-1.5B | A rented RTX 4500 Ada | A 2014-era GTX 1080 |
Every claim on these pages links back to a repository, and results are reported the way they came out — including ties, regressions and failed iterations.
Get involved
Help write the evals.
Two one-week, remote internships are open: curating evaluation prompts for Project Kestrel, or turning the research into something a curious non-expert can follow.